A Robust Object Detection System for Driverless Vehicles through Sensor Fusion and Artificial Intelligence Techniques

Abstract

Since the early 1990s, various research domains have been concerned with the concept of autonomous driving, leading to the widespread implementation of numerous advanced driver assistance features. However, fully automated vehicles have not yet been introduced to the market. The process of autonomous driving can be outlined through the following stages: environment perception, ego-vehicle localization, trajectory estimation, path planning, and vehicle control. Environment perception is partially based on computer vision algorithms that can detect and track surrounding objects. The process of objects detection performed by autonomous vehicles is considered challenging for several reasons, such as the presence of multiple dynamic objects in the same scene, interaction between objects, real-time speed requirements, and the presence of diverse weather conditions (e.g., rain, snow, fog, etc.). Although many studies have been conducted on objects detection performed by autonomous vehicles, it remains a challenging task, and improving the performance of object detection in diverse driving scenes is an ongoing field. This thesis aims to develop novel methods for the detection and 3D localization of surrounding dynamic objects in driving scenes in different rainy weather conditions.

In this thesis, firstly, owing to the frequent occurrence of rain and its negative effect on the performance of objects detection operation, a real-time lightweight deraining network is proposed; it works on single real-time images separately. Rain streaks and the accumulation of rain streaks introduce distinct visual degradation effects to captured images. The proposed deraining network effectively removes both rain streaks and accumulated rain streaks from images. It makes use of the progressive operation of two main stages: rain streaks removal and rain streaks accumulation removal. The rain streaks removal stage is based on a Residual Network (ResNet) to maintain real-time performance and avoid adding to the computational complexity.

Furthermore, the application of recursive computations involves the sharing of network parameters. Meanwhile, distant rain streaks accumulate and induce a distortion similar to fogging. Thus, it could be mitigated in a way similar to defogging. This stage relies on a transmission-guided lightweight network (TGL-Net). The proposed deraining network was evaluated on five datasets having synthetic rain of different properties and two other datasets with real rainy scenes.

Secondly, an emphasis has been put on proposing a novel sensory system that achieves realtime multiple dynamic objects detection in driving scenes. The proposed sensory system utilizes a monocular camera and a 2D Light Detection and Ranging (LiDAR) sensor in a complementary fusion approach. YOLOv3- a baseline real-time object detection algorithm has been used to detect and classify objects in images captured by the camera; detected objects are surrounded by bounding boxes to localize them within the frames. Since objects present in a driving scene are dynamic and usually occluding each other, an algorithm has been developed to differentiate objects whose bounding boxes are overlapping. Moreover, the locations of bounding boxes within frames (in pixels) are converted into real-world angular coordinates. A 2D LiDAR was used to obtain depth measurements while maintaining low computational requirements in order to save resources for other autonomous driving related operations. A novel technique has been developed and tested for processing and mapping 2D LiDAR measurements with corresponding bounding boxes. The detection accuracy of the proposed system was manually evaluated in different real-time scenarios. Finally, the effectiveness of the proposed deraining network was validated in terms of its impact on objects detection in the context of de-rained images.

Results of the proposed deraining network were compared to existing baseline deraining networks and have shown that the running time of the proposed network is 2.23× faster than the average running time of baseline deraining networks while achieving 1.2× improvement when tested on different synthetic datasets. Moreover, tests on the LiDAR measurements showed an average error of ±0.04m in real driving scenes. Also, both deraining and objects detection are jointly tested, and it was demonstrated that performing deraining ahead of objects detection caused 1.45× enhancement in the object detection precision.